我的 Gibbs 采样器中混合 model 的 NA 错误

Question

I am working on a Gibbs sampler and my code is as follows. The idea is (1)sample pi first (2) sample delta (3) sample beta .

  library(foreign)
   cognitive `=read.dta("http://www.stat.columbia.edu/~gelman/arm/examples/child.iq/kidiq.dta")`
 summary(cognitive)
 cognitive$mom_work = as.numeric(cognitive$mom_work > 1)
 cognitive$mom_hs = as.numeric(cognitive$mom_hs > 0)

# Modify column names of the data set
colnames(cognitive) = c("kid_score", "hs", "IQ", "work", "age")

 x<-cbind(cognitive$hs, cognitive$IQ, cognitive$work, cognitive$age)
 y<-cognitive$kid_score
 lmmodel<-lm(y~x-1, data=cognitive)
  NSim=3000 #iteration 
 Betahat=solve(t(x)%*%x)%*%t(x)%*%y

Error in if (delta[ite, j] == 1) rnorm(1, mu1, sigma1) else rnorm(1, mu0, : missing value where TRUE/FALSE needed In addition: Warning messages: 1: In rbinom(1, 1, prob = (p1/(p0 + p1))): NAs produced 2: In rbinom(1, 1, prob = (p1/(p0 + p1))): NAs produced

Answer 1

error is caused by the line prob=(pi[ite]*exp(-beta[ite-1,j]^2/(2*10^2)))/(((1-pi[ite])*10^3)*exp(-beta[ite-1,j]^2/(2*10^(-4)))+pi[ite-1]*exp(-beta[ite-1,j]^2/(2*10^2))) : at some iteration prob becomes greater than 1, so rbern() returns NA . Check your formula.

UPD. For debugging, add the following before your delta[ite,j]=rbern(... line:

prob_full <- (pi[ite]*exp(-beta[ite-1,j]^2/(2*10^2)))/(((1-pi[ite])*10^3)*exp(-beta[ite-1,j]^2/(2*10^(-4)))+pi[ite-1]*exp(-beta[ite-1,j]^2/(2*10^2)));
cat('\n',ite,j,prob_full)